210 likes | 251 Views
PDF uncertainties using a Monte Carlo method. Voica A. Radescu In collaboration with A. Glazov DESY Motivation Method Cross check of the method: Comparison with the standard error estimation Test various assumptions for the error distributions: Gauss, Log-normal, Uniform Summary.
E N D
PDF uncertainties using a Monte Carlo method Voica A. Radescu In collaboration with A. Glazov DESY Motivation Method Cross check of the method: Comparison with the standard error estimation Test various assumptions for the error distributions: Gauss, Log-normal, Uniform Summary
Motivation • The idea is to use a Monte Carlo method to estimate PDF uncertainties under various assumptions for the error distributions. • Standard error estimation of PDFs relies on the assumption that all errors follow Gauss statistics. Monte Carlo method can provide an independent cross check of it. • However, Gaussian assumption is not always correct: • Some systematic uncertainties follow Log-Normal Distribution: • lumi ,detector acceptance,etc. Gauss and Log-normal distributions: same mean shifted peaks How is that affecting PDF’s errors? • Some systematic uncertainties follow a Uniform Distribution: • “upper” limit uncertainties Monte Carlo method allows to test the various assumptions. Low X, Crete, Greece voica@mail.desy.de
Method (I) • Prepare “shifted data sets”: • Allow the central value of the cross sections ( ) to fluctuate within its systematic and statistical uncertainties taking into account all the correlations: • Various assumptions can be considered for the error distributions: • Gauss, Lognormal, Uniform Low X, Crete, Greece voica@mail.desy.de
Method (II) • Shifts for Statistical errors: • allow each data point to randomly fluctuate within its statistical uncertainty assuming either Gauss, Log-normal, or uniform distributions • Shifts for Systematic errors: • For each systematic source j uniformly select “fluctuation probability” Pj • For each data point shift the central value of cross section such that probability of this shift for systematic source j is equal Pj (or (1-Pj)) Low X, Crete, Greece voica@mail.desy.de
Method (III) • Repeat the preparations for N times (here N 100) • Perform the NLO QCD fit N times to extract PDFs • PDF uncertainties =>from the RMS of the spread • This study is performed using: • published H1-HERA I data of NC and CC e±p scattering cross sections [ref: Eur. Phys. J. C 30, 1-32 (2003)] • fit program H1 QCDNUM implementation at NLO: • MSbar renormalisation scheme, DGLAP evolution at NLO, massless quarks, polynomial form for PDF parametrisation a’ la H1PDF2000 Low X, Crete, Greece voica@mail.desy.de
Test of the Method: Gauss Distribution (xU) • To test the method, assume all errors follow Gauss distribution and compare the results to standard error estimation • 500 Green lines • Red lines: PDF uncertainties from RMS • Black lines: Hessian errors • Projections at various x values are shown: • Good Gauss Fit Good agreement Low X, Crete, Greece voica@mail.desy.de
Test of the Method: Gauss Distribution (xG) • To test the method, assume all errors follow Gauss distribution and compare the results to standard error estimation • 500 Green lines • Red lines: PDF uncertainties from RMS • Black lines: Hessian errors Good agreement with the standard error estimation! Good agreement Low X, Crete, Greece voica@mail.desy.de
Test various assumptions for the errors • Now that the method is cross checked, we are ready to test other assumptions: • Log-normal for lumi, all the rest set to Gauss • Log-normal for all systematic errors, Gauss for statistical uncertainty • Uniform for all errors Low X, Crete, Greece voica@mail.desy.de
1. Log-normal dist. for Lumi (xU) • Assume that all errors, apart from Lumi, follow Gauss • Test the effect of log-normal assumption for Lumi uncertainty • 100 Green lines • Red lines: PDF uncertainties from RMS • Black lines: Hessian errors Similar effect to pure gaussian case! Low X, Crete, Greece voica@mail.desy.de
1. Log-normal dist. for Lumi (xG) • Assume that all errors, apart from Lumi, follow Gauss • Test the effect of log-normal assumption for Lumi uncertainty • 100 Green lines • Red lines: PDF uncertainties from RMS • Black lines: Hessian errors Similar effect to pure gaussian case! Low X, Crete, Greece voica@mail.desy.de
2. Log-normal dist. for all systematic (xU) • Assume that statistical errors follow Gauss distribution and and all systematic errors follow Log-normal distribution • 100 Green lines • Red lines: PDF uncertainties from RMS • Black lines: Hessian errors Similar effect to pure gaussian case! Low X, Crete, Greece voica@mail.desy.de
2. Log-normal dist. for all systematic (xG) • Assume that statistical errors follow Gauss distribution, all systematic errors follow Log-normal distribution • 100 Green lines • Red lines: PDF uncertainties from RMS • Black lines: Hessian errors Similar effect to pure gaussian case! Low X, Crete, Greece voica@mail.desy.de
3. Uniform dist. for all errors (xU) • Assume all errors follow uniform distribution. • 100 Green lines • Red lines: PDF uncertainties from RMS • Black lines: Hessian errors Similar effect to pure gaussian case! Low X, Crete, Greece voica@mail.desy.de
3. Uniform dist. for all errors (xG) • Assume all errors follow uniform distribution. • 100 Green lines • Red lines: PDF uncertainties from RMS • Black lines: Hessian errors Similar effect to pure gaussian case! Low X, Crete, Greece voica@mail.desy.de
Summary • A simple method to estimate PDF uncertainties has been built within QCD Fit framework: • Assuming only Gaussian distribution of all errors, the results agree well with the standard error estimation • This method allows to check the effect of non-Gaussian assumptions for distributions of the experimental uncertainties: • For the H1 data, results are similar to the Gaussian case when using Log-normal and Uniform distributions of the uncertainties • The method could be extended for other physical variables (i.e. cross sections) for cross checks with the standard error evaluation Low X, Crete, Greece voica@mail.desy.de
Backup • Gauss Errors: more distributions Low X, Crete, Greece voica@mail.desy.de
Backup • Uniform Errors: more distributions Low X, Crete, Greece voica@mail.desy.de
Backup • All gauss, apart from Lumi (log-normal): Low X, Crete, Greece voica@mail.desy.de
Backup • Stat. gauss, sys log-normal: Low X, Crete, Greece voica@mail.desy.de
Method (IV): more details … • Use uniform random number to select the “fluctuation probability” Pj • Cross section point i has a sensitivity to systematic j • If > 0, select cross section shifts such that: • If > 0, select cross section shifts such that: Where P is either Gauss, Log-normal, or Uniform distribution Low X, Crete, Greece voica@mail.desy.de